Estimation of the past missing or unrecorded data (backward prediction) is very important, and it can be helpful in many applications. In this paper, backward prediction for multivariate ARMA processes is obtained by applying the forward representation for the processes which considerably simplifies the procedure.